DOORDASH FOR COUNTER TERRORIST
JTAC talks · AI parses · MCP tools verify · humans decide
LAYER 1 · OPERATOR (JTAC) — THE ONLY TRIGGER AUTHORITY
Natural language inputvoice or text — "troops in contact, grid 12345678…"
Brief confirmationapproves the AI-parsed four-input brief
"Cleared hot"human-only, never delegated, never automated
↓ plain speech in · validated brief back for confirmation
LAYER 2 · AI ORCHESTRATOR
Brief parserspeech → four inputs: location, description, mark, friendlies
Tool callerfans out to MCP tools, collects results
Discrepancy flaggerreadback diffs, geometry mismatches, danger-close
Readback composervalidated brief in JTAC's own format
↓ deterministic tool calls — the LLM never does math
LAYER 3 · MCP SERVER — THE TOOLBOX (CODE, NOT VIBES)
validate_gridformat + datum check
check_restrictedno-strike list, protected sites
compute_danger_closefriendlies vs. weapon effects
crosscheck_geometryIP/heading/distance consistency
get_terrain_elevationstated elevation vs. DTED
check_threat_ringsSAM/AAA on run-in and egress
suggest_weapondescription → pairing recommendation
transcribe_readbackvoice 9-line → diff vs. digital
log_missionimmutable record, every call
↓ results return up · nothing fires without passing Layer 1 again
LAYER 4 · HUMAN GATES (IN SERIES, NEVER PARALLEL)
Gate 1 — JTAC confirms briefAI's parse is a draft until a human signs it
Gate 2 — Pilot confirmssecond human, independent check
Gate 3 — Cleared hotthe only command that releases weapons
↓ everything, always
LAYER 5 · AUDIT LOG
Immutable mission recordevery input, tool call, gate decision, timestamped
Debrief replayreconstruct any decision after the fact
CHARGES — WHAT'S FREE, WHAT'S METERED
| TIER | WHAT'S IN IT | PRICE MODEL |
|---|
| SAFETY PATH — FREE, UNLIMITED |
validate_grid · check_restricted · compute_danger_close · crosscheck_geometry · get_terrain_elevation · check_threat_rings · transcribe_readback · log_mission |
Flat platform rate. Never per-call. Cost must never influence whether a check runs. |
| PREMIUM — METERED |
suggest_weapon tuning · demand forecasting · multi-mission planning · 3D terrain visualization · debrief analytics |
Pay-per-use or credits up front. Meter the nice-to-haves, never the trigger-pull. |
| ENTERPRISE |
Dedicated MCP hosting · custom tool development · on-prem deployment · SLA |
Annual contract. |
DESIGN PRINCIPLES
- LLM for fuzzy, code for math. The orchestrator parses speech; the MCP tools do every calculation. A language model never computes danger-close.
- Never meter the safety path. Per-call pricing on validation creates a financial incentive to skip checks. The safety tier is flat-rate and unlimited by design, not by generosity.
- Humans own every trigger decision. Three gates in series. The AI drafts, validates, and flags — it never clears hot.
- Parsing a wrong grid faster is still a wrong grid. The product's value is cross-validation catching the misheard digit, not the stopwatch.
- Everything is logged. If a decision can't be reconstructed afterward, the system failed regardless of the outcome.